What does it feel like to ship something almost entirely on your own?

Every launch I've been part of before this one was loud. A team, a war room, a countdown, everyone refreshing the same dashboard.

Kakashi is out now, and it was a different kind of launch. Quiet. Mostly just me.

It watches your jiu-jitsu rolls and tells you what actually happened - the positions, the transitions, where things went sideways. We ran a small alpha with a few gyms and athletes first, and the feedback was really positive, more than I expected honestly. That's what gave me the nerve to open it up.

When I say I built nearly all of it myself, I mean it in the most unglamorous way possible. The analysis pipeline that watches the footage, the app people upload to, the payments so it can actually pay for its own compute. A year ago that would've needed a team. Today it needs one stubborn person, a stack of AI subscriptions, and a willingness to stay up far too late. I still design the solution and the architecture, and the models write a lot of the code with me in the loop, reviewing, rejecting, redirecting. It's made shipping alone possible, not easy.

But here's the part I didn't see coming.

The most valuable thing that happened during the alpha wasn't anyone saying they liked it. It was coaches going through the outputs and correcting them. The kind of note that says this was a different guard, or this transition happened earlier, or you've called two different things by the same name. At first every correction felt like a small punch to the gut, because I'd built this thing and I wanted it to be right. Then it clicked that this was the product working. Every correction a coach makes is a piece of knowledge I couldn't have written down myself, flowing back into the system and making it a little less wrong for the next person.

I've always said I can't do it alone, and I kept saying it right up until I was building something alone. Turns out I was never actually alone. The coaches are the team. They just don't have commit access.

The other thing the alpha taught me is that this is a long-horizon problem, in both senses. Technically, the hardest part of understanding a roll is time - remembering what happened three minutes ago so you can make sense of what's happening now, when two bodies are wrapped around each other and half the limbs are hidden. The models are brilliant for a few seconds and then quietly lose the plot. And as a project, this isn't something that gets solved in a launch. It's something that gets a little better every week as more rolls come in and more coaches push back on what it says.

I used to be more impatient about this kind of thing. The mats cured me of some of that. You don't get good at jiu-jitsu in a season. You show up, you get smashed, you learn one small thing, you come back tomorrow.

So no big numbers to share, no screenshot of a dashboard going vertical. Just a thing that's out in the world, a handful of gyms who've been generous with their time, and a very long list of corrections I'm grateful for. It's buggy in places, it gets confused when things get tangled, and it's already more honest than it was in January.

I went to bed exhausted the night it launched, knowing I'd given it my all. Then I woke up and read the first round of corrections.

This is going to take years. Good. I'm not going anywhere.